CountEx: Fine-Grained Counting via Exemplars and Exclusion

Yifeng Huang; Gia Khanh Nguyen; Minh Hoai

cs.CVFeb 23, 2026

CountEx: Fine-Grained Counting via Exemplars and Exclusion

Authors: Yifeng Huang, Gia Khanh Nguyen, Minh Hoai

Organizations: Department of Computer Science, Stony Brook University, Stony Brook, NY, USA · Australian Institute for Machine Learning, Adelaide University, SA, Australia

Abstract

This paper presents CountEx, a discriminative visual counting framework designed to address a key limitation of existing prompt-based methods: the inability to explicitly exclude visually similar distractors. While current approaches allow users to specify what to count via inclusion prompts, they often struggle in cluttered scenes with confusable object categories, leading to ambiguity and overcounting. CountEx enables users to express both inclusion and exclusion intent, specifying what to count and what to ignore, through multimodal prompts including natural language descriptions and optional visual exemplars. At the core of CountEx is a novel Discriminative Query Refinement module, which jointly reasons over inclusion and exclusion cues by first identifying shared visual features, then isolating exclusion-specific patterns, and finally applying selective suppression to refine the counting query. To support systematic evaluation of fine-grained counting methods, we introduce CoCount, a benchmark comprising 1,780 videos and 10,086 annotated frames across 97 category pairs. Experiments show that CountEx achieves substantial improvements over state-of-the-art methods for counting objects from both known and novel categories. The data and code are available at https://github.com/bbvisual/CountEx.

Figures & tables

Figure 1: Given a cluttered scene containing mutiple object categories, our method allows users to specify both inclusion and exclusion intent, e.g., “Count penne pasta, not spiral pasta”, via language prompts and optional visual exemplars, enabling more precise and controllable counting.

Figure 2: Object categories and their variants across five super-categories on CoCount.

		Train	Train2	Val	Test
# videos	Food	92	92	92	92
	Game	92	92	92	92
	Home	95	95	95	95
	Desk	86	86	86	86
	Misc	80	80	80	80
	Total	445	445	445	445

Table 1: Video counts and annotation across dataset splits.

Figure 3: Overview of CountEx. Given an image and multimodal prompts (positive and negative text with optional visual exemplars), we encode them into query sets $\mathbf{Q}^{\text{pos}}$ and $\mathbf{Q}^{\text{neg}}$ . Our Discriminative Query Refinement (DQR) module consists of three stages: (1) Shared Feature Identification learns prototypes $\mathbf{C}$ capturing common features between both query sets; (2) Exclusive Feature Extraction isolates negative-exclusive patterns $\mathbf{R}^{\text{neg}}$ by projecting $\mathbf{Q}^{\text{neg}}$ onto $\mathbf{C}$ and filtering residuals; (3) Selective Query Refinement produces refined queries $\tilde{\mathbf{Q}}^{\text{pos}}$ by selectively suppressing negative patterns via attention. The refined queries are fed to detection heads for final predictions. An auxiliary density prediction branch provides additional supervision during training.

Method	NC-setting		KC-setting
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	33.22	47.66	16.82	29.23
CAD-GD [ 44 ]	34.08	50.39	16.00	27.52
GroundingREC [ 5 ]	29.29	42.43	17.54	27.41
CountGD [ 2 ]	33.78	48.29	15.55	28.32
CountEx (proposed)	26.61	38.86	12.72	23.99

Table 2: Performance comparison on the CoCount test set. All methods are trained using the same data under either the known-category (KC) or novel-category (NC) setting. CountEx outperforms all baselines across both settings.

(per category synthetic data generation + adaptation)
Method	MAE	RMSE
D’Alessandro et al. [ 8 ]	10.00	16.95
Zero-shot - no training/adaptation
OWLv2 [ 30 ]	37.25	55.01
GroundingDINO [ 24 ]	33.89	59.51
PseCo [ 13 ]	59.82	72.79

Table 3: Results on LOOKALIKES [ 8 ] . * denotes training on CoCount before zero-shot evaluation. CountEx achieves best zero-shot performance. D’Alessandro et al. [ 8 ] obtain lowest MAE but require synthetic data generation and per-category test-time adaptation (5–7 minutes per category).

Pre-trained specialist counting models
Model	MAE $\downarrow$		NAE $\downarrow$
	Inter	Intra	Inter	Intra
DAVE [ 34 ]	46.27	46.75	0.779	0.797
GeCo [ 33 ]	45.05	54.80	0.777	0.935
LoCA [ 6 ]	71.89	57.45	1.177	0.950
FamNet [ 38 ]	66.97	74.75	1.363	1.440

Table 4: Results on PairTally [ 31 ] . CountEx achieves the best performance on both MAE and NAE, outperforming pre-trained specialist counters, general vision-language models, and models trained on the same CoCount dataset.

Input Modalities				NC-setting		KC-setting
$T_{\text{pos}}$	$E_{\text{pos}}$	$T_{\text{neg}}$	$E_{\text{neg}}$	MAE	RMSE	MAE	RMSE
✓				32.22	48.38	15.96	28.76
✓	✓			31.57	46.73	15.75	27.59
✓		✓		26.67	39.00	13.22	26.23
✓	✓	✓	✓	26.61	38.86	12.72	23.99

Table 5: Performance of different combinations of multimodal prompts on the test set of CoCount.

$\mathcal{L}_{\text{den}}$	$\mathcal{L}_{\text{proto}}$	NC-setting		KC-setting
$\mathcal{L}_{\text{den}}$	$\mathcal{L}_{\text{proto}}$	MAE	RMSE	MAE	RMSE
		30.75	42.84	15.20	28.33
✓		29.21	41.61	13.00	25.22
✓	✓	26.61	38.86	12.72	23.99

Table 6: Ablation on loss components. Both density prediction loss (

\mathcal{L}_{\text{den}}

) and the proto losses (

\mathcal{L}_{\text{proto}}

) contribute to improved performance.

Panel of Figure 4 — Figure 4: Qualitative results on CoCount test set.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	17.60	30.50	16.82	29.23
CAD-GD [ 44 ]	15.85	26.92	16.00	27.52
GroundingREC [ 5 ]	17.72	27.83	17.54	27.41
CountGD [ 2 ]	16.06	28.39	15.55	28.32
CountEx (proposed)	14.07	27.61	12.72	23.99

Table 7: Complete results on CoCount KC-setting.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	45.87	65.71	45.92	65.69
CAD-GD [ 44 ]	56.52	81.25	56.94	81.64
GroundingREC [ 5 ]	41.24	58.28	40.85	57.22
CountGD [ 2 ]	44.00	60.17	41.63	57.50
CountEx (proposed)	39.41	54.82	37.04	50.58

Table 8: Complete results on NC-setting Test-Food.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	27.75	41.64	27.77	41.61
CAD-GD [ 44 ]	24.57	37.05	26.36	39.96
GroundingREC [ 5 ]	28.26	41.80	29.31	43.55
CountGD [ 2 ]	35.06	48.92	32.73	47.24
CountEx (proposed)	25.55	37.72	24.16	34.87

Table 9: Complete results on NC-setting Test-Home.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	35.69	57.57	35.72	57.55
CAD-GD [ 44 ]	39.28	57.29	36.28	54.47
GroundingREC [ 5 ]	36.27	54.56	33.31	51.24
CountGD [ 2 ]	41.01	54.69	39.40	58.08
CountEx (proposed)	32.07	49.14	31.18	51.90

Table 10: Complete results on NC-setting Test-Desk.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	32.51	42.56	32.39	42.44
CAD-GD [ 44 ]	30.48	47.82	32.13	48.84
GroundingREC [ 5 ]	26.49	39.80	25.04	34.38
CountGD [ 2 ]	28.87	38.59	31.96	42.97
CountEx (proposed)	26.78	37.91	23.82	32.68

Table 11: Complete results on NC-setting Test-Misc.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
LLMDet [ 9 ]	23.66	30.17	24.31	31.00
CAD-GD [ 44 ]	19.64	28.10	18.67	27.02
GroundingREC [ 5 ]	21.86	31.72	17.94	25.75
CountGD [ 2 ]	20.92	31.44	23.18	35.68
CountEx (proposed)	16.64	23.50	16.84	24.26

Table 12: Complete results on NC-setting Test-Game.

Method	Val		Test
Method	MAE	RMSE	MAE	RMSE
CounTR [ 21 ]	13.13	49.83	11.95	91.23
LOCA [ 6 ]	10.24	32.56	10.79	56.97
CACViT [ 45 ]	10.63	37.95	9.13	48.96
DAVE [ 34 ]	8.91	28.08	8.66	32.36
CountGD [ 2 ]	7.10	26.08	5.74	24.09

Table 13: Complete results on FSC-147.

Figure 6: Object count distribution per image.

Figure 7: Additional qualitative results on CoCount test set (KC-setting).

Panel of Figure 8 — Figure 8: Additional qualitative results on CoCount test set (NC-setting).

Object Counting

Object counting estimates the number and locations of target instances in images or video, from dense crowds to fine-grained categories.

See all 58 papers

Count Anything at Any Granularity
May 11, 2026Chang Liu, Haoning Wu, Weidi XieObject Counting Multimodal Large Language Models
School of Artificial Intelligence, Shanghai Jiao Tong University, China · CMIC, Shanghai Jiao Tong University, China
RECOUNT: Reference-guided Counting with Synthetic Visual Exemplars
Aug 20, 2026Adriano D'Alessandro, Ali Mahdavi-Amiri, Ghassan HamarnehObject Counting
Simon Fraser University
Count Anything
May 29, 2026Mengqi Lei, Shuokun Cheng, Wei Bao +4Object Counting
1{BNRist, THUIBCS, BLBCI, School of Software}, Tsinghua University · School of Computer Science and Technology, China University of Geosciences, Wuhan · 3State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University

May 11, 2026J/K move · Enter open · S save

Count Anything at Any Granularity

Chang Liu, Haoning Wu, Weidi Xie

School of Artificial Intelligence, Shanghai Jiao Tong University, China · CMIC, Shanghai Jiao Tong University, China

Object Counting Multimodal Large Language Models Fine-Grained Visual Perception Multimodal Understanding

Open paper AskPDF arXiv

Open-world object counting remains brittle: despite rapid advances in vision-language models (VLMs), reliably counting the objects a user intends is far from solved. We argue that a central reason is that counting granularity is left implicit; users may refer to a specific identity, an attribute, an instance type, a category, or an abstract concept, yet most methods treat "what to count" as a single, category-level matching problem. In this work, we redefine open-world counting as multi-grained counting, where visual exemplars specify target appearance and fine-grained text, with optional negative prompts, specifies the intended semantic granularity across five explicit levels. Making granularity explicit, however, exposes a critical data bottleneck: existing counting datasets lack the multi-category scenes, controlled distractors, and instance-level annotations needed to verify fine-grained prompt semantics. To address this, we propose the first fully automatic data-scaling pipeline that integrates controllable 3D synthesis with consistent image editing and VLM-based filtering, and use it to construct KubriCount, the largest and most comprehensively annotated counting dataset to date, supporting both training and multi-grained evaluation. Systematic benchmarking reveals that both multimodal large language models and specialist counting models exhibit severe prompt-following failures under fine-grained distinctions. Motivated by these findings, we train HieraCount, a multi-grained counting model that jointly leverages text and visual exemplars as complementary target specifications. HieraCount substantially improves multi-grained counting accuracy and generalizes robustly to challenging real-world scenarios. The project page is available here: https://verg-avesta.github.io/KubriCount/.

Abstract

Object Counting

Explore similar work

Count Anything at Any Granularity

RECOUNT: Reference-guided Counting with Synthetic Visual Exemplars

Count Anything